Image skeleton is a compact and intuitive image representation method.Currently,the existing skeleton ex-traction of binary pixel images based on depth learning faces the problem of skeleton breakage.This paper proposes a fusing full-scale side outputs residual unet network(FFSR-Unet)of skeleton extraction algorithm.This network achieves fea-ture interaction of foreground objects with different shape scales by fusing features between different levels of encoders and decoders,and uses Stepwise-Resblock to enhance the network's ability to extract deep and shallow semantics.On the Pixel SkelNetOn Challenge dataset,the F1-score obtained according to this network model can reach 0.854 8,which is able to surpass the extraction results of existing algorithms.
维普
万方数据